Abstract

The advance of misinformation on digital platforms has emerged as a major social affair, influencing public opinion, destabilizing democratic processes, and undermining trust in institutions. Traditional detection methods, which rely on rule-based logic and keyword matching, have proven insufficient in addressing the emerging complexity of fraudulent content. Although deep learning models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) perform better, they continue to face limitations in scalability, generalization, and understandability. This paper presents a comprehensive framework for misinformation detection that uses the capabilities of transformer-based architectures, in particular BERT (Bidirectional Encoder Representations from Transformers), enhanced with explainable artificial intelligence (XAI) techniques including SHAP. These methods contribute to the understandability of model decisions, encourage transparency, and foster user trust. To further address the challenge of limited labeled datasets and domain-specific differences in misinformation, a semi-supervised learning policy is included, enabling the model to leverage both labeled and unlabeled data for better generalization across different disinformation domains. Using the political entity subset of the FakeNewsNet dataset, a comprehensive performance evaluation of the BERT-based structure was conducted against classical machine learning baselines such as Naive Bayes and Support Vector Machines (SVM). Experimental results show that the proposed model notably outperforms traditional methods, achieving high accuracy while providing understandable results, and demonstrating strong predictive power alongside the interpretability needed for real-world deployment in fighting disinformation.